For sixty years, the structural unit of primary care has been the visit. A 20-minute appointment. A scheduled slot. A clinician who walks into a room, opens a chart from scratch, takes a history, makes a decision, writes a note, and moves on. Eight hours, twenty-four visits, roughly twelve hundred patients in panel.
That's the math. It hasn't moved in two generations.
When the country asks how do we scale healthcare?, the answer has always been the same: add clinicians. Hire more physicians. Train more nurse practitioners. Build more clinics. Each new clinician adds another twenty-four visits a day to the total. The structure is linear; the workforce grows in lockstep with the patient population.
It hasn't worked. US medical schools graduate roughly thirty thousand MDs a year. Roughly ten thousand of them enter primary care. We have around 120 million American adults living with chronic disease that requires longitudinal management. Staffing the chronic care population on the visit-based model would consume a decade of US primary care graduates — and that's just to keep up, not to grow. The model does not fit the workforce. It never has.
The instinct is to demand we hire faster. The actual question is: what if the unit of work weren't a 20-minute visit?
The traditional MD's day is the constraint
Watch one MD's day in a traditional clinic or its telemedicine equivalent. They start at 8 am. They have twenty-four slots, each twenty minutes long. They see Patient 1 at 8:00, Patient 2 at 8:20, and so on through 4 pm. Every clinical decision in their panel runs through them: every refill, every blood pressure follow-up, every routine titration, every patient question — all twenty minutes long, all built from scratch.
The day is wall-to-wall. The MD's direct-care share — the fraction of clinical decisions in the panel that personally require their reasoning — is 100%.
Now add a midlevel: a nurse practitioner running visits in a parallel room. The MD does sixteen of those twenty-minute visits; the NP does sixteen. Combined, the team handles thirty-two visits a day, and the panel grows to about 2,000. Better. But notice what didn't change. The MD's clock is still 100% full — the visits the MD doesn't do, they fill with admin, inbox, and coordination of the NP's work. The MD's direct-care share drops to 50%, but only because the NP picked up half the visits, not because the MD has slack.
This is the secret of every visit-based staffing model. Adding a midlevel doesn't free MD time. It adds another room running in parallel. The output goes up; the bottleneck doesn't move.
The figure below makes the contrast visible. Each row is one MD's eight-hour day under a different staffing model. Press Play the day and watch the clock advance from 8 am to 4 pm. The strips fill in real time with the MD's actual clinical interactions. Visit-based strips fill wall-to-wall. Actually Health strips stay mostly empty.
Interactive figure
Same hours. The MD doesn't shrink — the population the MD reaches grows.
Each row is one MD's eight-hour workday. The gauge below each strip is the MD's direct-care share — the fraction of clinical decisions that personally require the MD's reasoning. Press "Play the day" to advance the clock 8a → 4p. In visit-based models the MD's share is 100% (every decision is theirs) — and the panel maxes at 1,200–2,000. In Actually Health's architecture, the MD's share is designed to fall to 15% (or 5% with a midlevel) — and the model targets a panel of 3,500–5,500. Pre-launch design; not measured.
⚡ The architectural bet: the MD becomes leveraged, not busier
The traditional MD personally closes 100% of clinical decisions for a 1,200-patient panel. Adding a midlevel cuts MD share to 50% but only doubles the panel — still visit-bound. Switch to async + AI distillation: most clinical loops arrive pre-assembled so the MD verifies in seconds. By design, MD share falls to ~15% and the panel target rises to ~3,500. Add a midlevel on top — the NP/PA closes routine bundles, leaving the MD on diagnostic uncertainty, multi-problem conflict, drug interactions, and safety escalations. MD share is targeted at ~5%; panel target reaches ~5,500. The bet of the architecture is that the same MD-hour budget reaches more of the population by reaching less of it directly. (Pre-launch projection from the model.)
What async + AI distillation actually does
The Actually Health rows on the figure show something visit-based care has no slot for. Most of the day is empty. Scattered across the eight hours are short navy bars and thin gold ticks. The navy bars are 4-minute decision bundles: structured clinical reasoning where the system has pre-assembled the case — the typed signals, the options ruled in and out with reasoning, the cited guideline criteria, the recommended action, the follow-up conditions. The MD verifies, modifies, or signs. The gold ticks are 30-second AI-distilled verifications: a home BP reading in range, a refill within rule, a structured education question. The MD glances and signs off.
This is the architectural unlock. Visit-based care charges full price — twenty minutes — for every clinical question, routine or complex. Async care charges by complexity. Routine: thirty seconds. Standard: four minutes. Genuinely hard: still four to ten minutes, but on a pre-assembled foundation rather than rebuilt from scratch.
AI does not prescribe. AI distills and presents; the MD remains the clinical decision-maker on every loop. What changes is the cost of each loop. Drop the cost from twenty minutes to four — or to thirty seconds — and the MD's eight-hour day stretches to cover three times the panel.
That's what the third row of the figure projects: a panel target of ~3,500 patients per MD, with MD direct-care share designed to drop to about 15%. Same eight hours. Same two-person team. Different unit of work. This is the structural target — we're building toward it, not reporting it.
With a midlevel, the MD becomes the specialist of the team
Add a nurse practitioner or physician assistant to the Actually Health pod and a second compounding lever kicks in. The NP/PA absorbs the routine clinical bundles — the protocol-bound titrations, the stable-patient reassessments, the in-range monitoring with rule-based responses. What's left for the MD is the work that doesn't fit a protocol.
This is what the fourth row of the figure models. Six bundles. Fifteen verifications. A panel target of ~5,500 patients per MD. The MD's direct-care share is designed to fall to ~5%. The MD is reserved for:
- Diagnostic uncertainty — novel or atypical presentations the system can't classify
- Multi-problem conflict — when two or more conditions interact and the right move requires synthesis
- Drug interaction review — regimen changes that span multiple problems
- Safety escalations — red flags, threshold breaches, deterioration signals
- Protocol exceptions — cases the NP/PA cannot close within rule
- Provisional → working diagnosis transitions — when the clinical picture is still resolving
Everything else flows around the MD. Routine refills go to AI distillation and a glance-verify. In-range BP confirmations go to AI distillation. Education delivery goes to the Navigator. Scheduling and structured intake go to the Navigator. Protocol-bound clinical titrations go to the NP. The MD's eight hours fill with the cases that genuinely require their training.
The doctor is reserved, not removed
Quality concerns sit on this distinction. If the MD signs off on five percent of clinical decisions directly, who is responsible for the other ninety-five? The MD still is — for protocol governance, for escalation pathways, for the cases that reach them, for the structure that determines which cases reach them. The architectural design says: the system surfaces the cases that need a doctor's reasoning, and the doctor reasons. The rest is closed by the appropriate role at the appropriate tier.
Burnout concerns sit on the same distinction. The traditional MD's day is a procession of routine encounters interleaved with hard ones, each consuming the same twenty minutes, each requiring chart reconstruction. The cognitive load is sustained at high level for eight hours regardless of complexity. The Actually Health MD's day is a smaller volume of harder, more interesting cases — exactly the work the doctor trained for. The routine work is absorbed below them. The administrative work is absorbed elsewhere. What remains is judgment.
Same hours. Same doctor. More patients reached. Less burnout exposure. The math is structural — and once the architecture is built for it, every additional patient absorbed comes with a smaller marginal cost.
The leapfrog looks small at one doctor. It's the only thing that matters at a country.
Per MD, the architecture takes the panel target from 1,200 to 5,500 — a projected 4.6× lift. That number sounds incremental until you scale it.
At one million patients, the visit-based model needs roughly 833 MDs dedicated to chronic care. The full Actually Health pod, if the panel targets hold, would need roughly 182. The implied gap is six hundred fifty MDs — a workforce delta larger than most regional health systems can hire in a decade.
At one hundred million patients — the US adult chronic disease population — the implied gap reaches tens of thousands of MDs. More clinicians than the country can produce in primary care across a decade. The visit-based model does not fit the workforce we have or the workforce we can train. An asynchronous architecture is built to — because each additional patient absorbed would cost the MD only the fraction of their time that they alone can supply: the judgment.
That is the leapfrog the architecture is designed to deliver. It looks like a per-doctor optimization at the basic unit. If the targets hold, it is the difference between deliverable chronic care and a shortage we cannot hire our way out of.